Researchers from various institutions have conducted an industrial case study on predicting the severity of software vulnerabilities using in-context learning. They used locally deployable, open-source large language models (LLMs) to directly predict Common Vulnerability Scoring System (CVSS) scores from vulnerable C/C++ code snippets. The team compared their results with a proprietary dataset and found that an open-source dataset called Big-Vul can be used as a proxy for industrial data when constructing testbeds. They also evaluated the performance of different LLM models, including CodeLlama2-7B, which showed promising results in approximating cloud performance for CVSS regression when guided by lightweight prompts. --- Why it matters: This research matters to AI engineers and researchers because it provides a practical solution for severity triage in industrial settings, reducing reliance on cloud-based services that raise confidentiality concerns. The study's findings can help improve the scalability of vulnerability assessment and reduce exposure to high-impact security flaws. Source: https://arxiv.org/abs/2608.22089